The Approval Gradient: Why LLMs Fail as Thinking Partners, and What Actually Works
We assume sycophancy in AI is a politeness problem, solvable with a sharper prompt that demands directness and tells the model to disregard our feelings. It isn’t. The sharper prompt produces a different version of the same failure. What actually works is more counterintuitive, and it changes how we should think about working with LLMs on serious intellectual work: software architecture, requirements analysis, incident reviews, governance decisions, and any situation where the cost of a politely reinforced false premise is high. ...